[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-baidu-wenxin-search-to-free-agent-en":3,"article-related-baidu-wenxin-search-to-free-agent-en":29,"series-tools-d5ff5ff4-6bed-4820-8ce4-650d9a5a6eb5":74},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":11},"d5ff5ff4-6bed-4820-8ce4-650d9a5a6eb5","baidu-wenxin-search-to-free-agent-en","Baidu Wenxin turns search into a free agent","\u003Cp data-speakable=\"summary\">Baidu Wenxin turns search-style retrieval into a free \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa> that finishes tasks.\u003C\u002Fp>\u003Cp>I've been watching agent products for a while now, and honestly, a lot of them still feel like fancy autocomplete with a tool belt. They can talk a good game, they can even sound confident while they work, but the second you ask them to hold context, pull files together, or finish something messy end to end, they start wobbling. That part is annoying because the pitch is always the same: “it can do work for you.” Sure. But can it actually deliver a usable artifact without me babysitting every step? That's the bar. So when I saw \u003Ca href=\"https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2068369583216537702\">this Zhihu post\u003C\u002Fa> about Baidu Wenxin Assistant apparently taking first place in both Chinese and global agent benchmarks, and doing it for free, I paid attention. Not because leaderboard posts usually change my mind, but because the shape of the claim matched the thing I keep running into in real projects: the hard part is not chatting, it's execution.\u003C\u002Fp>\u003Cp>The source that pushed me to break this down is a post on Zhihu by 新智元, published at \u003Ca href=\"https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2068369583216537702\">zhuanlan.zhihu.com\u002Fp\u002F2068369583216537702\u003C\u002Fa>. The article claims Wenxin Assistant scored 97.62 on SuperCLUE XClaw and 94.6% on PinchBench v2, while staying free and unlimited. I am not independently verifying those numbers here; I'm using the post as a case study for how a search company is trying to turn old retrieval muscle into agent behavior.\u003C\u002Fp>\u003Ch2>It is not about answering better. It is about finishing the job.\u003C\u002Fh2>\u003Cblockquote>“衡量一个AI好不好，标准正在从‘答得对不对’换成‘事办没办成’。”\u003C\u002Fblockquote>\u003Cp>What this actually means is simple: I do not care if the model sounds smart if I still have to do the last mile myself. The article keeps hammering this point, and I think it's the right one. Traditional chat models optimize for response quality. Agent systems optimize for task completion. That sounds like a small wording change, but in practice it flips the product from “conversation” to “output.”\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786365227346-puo3.png\" alt=\"Baidu Wenxin turns search into a free agent\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>I ran into this exact problem building internal assistants for research and operations. The model would summarize a document beautifully, then fail to produce the spreadsheet, the slide deck, or the formatted report the team actually needed. Everyone would nod at the answer and then open three more tools. That is not automation. That is a polite detour.\u003C\u002Fp>\u003Cp>The Zhihu piece frames Wenxin Assistant as a system that can read, plan, call tools, and return something usable. That matters more than raw chat fluency. If the agent can produce a Word document, a markdown report, or a runnable HTML artifact, then the user gets something they can hand off, ship, or review. If not, the whole “agent” label is just marketing wallpaper.\u003C\u002Fp>\u003Cp>How to apply it: when you evaluate an agent, stop asking whether it is “smart.” Ask whether it can complete one ugly, multi-step task without you intervening. Give it a file, a target format, a deadline, and a failure condition. Then watch whether it finishes or just narrates.\u003C\u002Fp>\u003Cul>\u003Cli>Test end-to-end delivery, not just intermediate reasoning.\u003C\u002Fli>\u003Cli>Measure whether the output is directly usable by a human or downstream system.\u003C\u002Fli>\u003Cli>Track how many times you have to step in and patch the result.\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>XClaw matters because it punishes hand-wavy demos\u003C\u002Fh2>\u003Cp>The article points to SuperCLUE’s XClaw \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> as the Chinese-side proof. I am more interested in the benchmark design than the leaderboard drama. According to the post, XClaw tests five dimensions: data processing, content creation, memory, code, and research analysis. It also says the tasks are delivery-based, not multiple-choice. That is the part I care about.\u003C\u002Fp>\u003Cpre>\u003Ccode>“不要选择题，只看交付题。”\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>What this actually means is that the benchmark is trying to block the usual escape hatches. A model cannot bluff its way through a choice list. It has to produce the thing. That changes the game completely, because now you are testing whether the system can follow instructions, hold constraints, and assemble an artifact that survives inspection.\u003C\u002Fp>\u003Cp>I like this style of evaluation because it looks more like real work. Nobody in a company asks an agent to pick option B from a list. They ask it to clean up a dataset, write the brief, draft the plan, or generate the page. If your benchmark does not resemble that, I do not trust the score much.\u003C\u002Fp>\u003Cp>The article says Wenxin’s strongest XClaw result was memory, where it got 100. Then data processing at 98.86, content creation at 99.44, and research analysis at 96.44. I am not treating those numbers as universal truth here, but I do think the pattern is useful. Memory and research are exactly where agents usually fall apart. They forget constraints, lose thread, and start freelancing. If a system is strong there, it is probably doing more than stringing together canned responses.\u003C\u002Fp>\u003Cp>How to apply it: build your own benchmark around artifacts. Use tasks like “turn this brief into a Word doc,” “summarize these five files into one memo,” or “research this topic and cite sources.” Then score the final artifact, not the chat transcript.\u003C\u002Fp>\u003Cul>\u003Cli>Keep the task closed enough that you can judge success quickly.\u003C\u002Fli>\u003Cli>Include at least one constraint the model cannot ignore, like format or length.\u003C\u002Fli>\u003Cli>Penalize hallucinated structure, not just factual errors.\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>Memory is the boring feature that decides whether the agent is useful\u003C\u002Fh2>\u003Cp>The article spends a lot of time on memory, and that is fair. Memory is not sexy, but it is the thing that keeps an agent from acting like it has amnesia every third turn. The post describes memory as the foundation of AI work ability, and I agree with that more than I want to. If the system cannot remember what you said three messages ago, it is not helping. It is just consuming tokens.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786365229594-amwx.png\" alt=\"Baidu Wenxin turns search into a free agent\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>I have seen this in long-running support and ops flows. A user gives three conditions, the model acknowledges all three, and then the final output quietly forgets one. That is the kind of failure that makes teams stop trusting the tool. Not because the model was wildly wrong, but because it was inconsistently right.\u003C\u002Fp>\u003Cp>The Zhihu piece ties Wenxin’s memory strength to its search-engine background, which is actually a decent explanation. Search systems already live in a world of session context, query intent, and result ranking. Agent memory is not identical, but the shape is close enough that the engineering instincts transfer. If you have spent twenty years building systems that interpret user intent and preserve context across interactions, you are not starting from zero when you build an agent.\u003C\u002Fp>\u003Cp>How to apply it: treat memory as a product feature, not a model bonus. Decide what should persist across turns, what should expire, and what the user can explicitly pin. Then test those rules with annoying cases, not clean demos.\u003C\u002Fp>\u003Cp>One more thing: memory should be visible. If the agent is carrying context forward, I want to know what it thinks it knows. Hidden memory is how you get weird, untraceable failures later.\u003C\u002Fp>\u003Ch2>Data processing is where fake confidence dies fast\u003C\u002Fh2>\u003Cp>The article says Wenxin scored 98.86 in data processing, and that is the kind of category where I immediately ask, “Okay, but did it actually respect the numbers?” Data work is unforgiving. One wrong field, one sloppy rounding decision, one invented total, and the whole output is trash.\u003C\u002Fp>\u003Cp>In the post’s example, the model takes uploaded company material, extracts it with pandoc, structures it, and generates a formatted Word document. The claimed result is a 6,000-plus-word business plan with 148 sections. That kind of workflow is exactly what agents should be doing: ingest, structure, format, hand off. No drama, no “here’s a draft outline if you want.”\u003C\u002Fp>\u003Cp>What this actually means is that the agent is not just writing text. It is operating on documents as objects. That is a different skill. It requires parsing, preserving structure, and outputting in a file format that survives real use. If you have ever had to clean up a model-generated report that ignored table boundaries or mangled headings, you know why this matters.\u003C\u002Fp>\u003Cp>I ran into this when I tried to automate internal planning docs. The model could write a decent narrative, but it kept inventing section breaks and flattening tables. Once I forced it into a document pipeline with explicit extraction and formatting steps, the output got much better. Not because the model got more poetic, but because the workflow got stricter.\u003C\u002Fp>\u003Cp>How to apply it: if your use case involves spreadsheets, reports, or structured docs, do not start by asking for prose. Start by defining the schema, then the transformation, then the export format. Make the agent prove it can preserve structure before you let it write freely.\u003C\u002Fp>\u003Ch2>Research analysis is just search with better manners\u003C\u002Fh2>\u003Cp>The article’s strongest argument, at least to me, is that research analysis is where Baidu’s background pays off. The post describes a workflow where Wenxin plans a six-step research process, calls multiple sub-agents in parallel, searches academic sources, cross-checks claims, and then outputs both markdown and interactive HTML. That is not a chat trick. That is a research pipeline.\u003C\u002Fp>\u003Cp>One example in the post uses the Three-dimensional Kakeya conjecture and another builds a report on recent global model pricing and capability comparisons. In both cases, the key move is not that the agent “knows” the topic. The key move is that it can search, compare, verify, and then assemble a report that looks like something a human analyst would actually send around.\u003C\u002Fp>\u003Cp>This is where the search-company angle becomes interesting. Search engines already know how to break a query into intent, retrieve sources, rank results, and return a synthesized answer. Agents extend that by adding tool use and longer execution chains. So when the article says Baidu’s twenty years of search work became the backbone of Wenxin’s agent behavior, that sounds plausible to me. It is not magic. It is a migration path.\u003C\u002Fp>\u003Cp>How to apply it: if you are building research agents, do not let them jump straight to conclusions. Force a plan, force source collection, force cross-checking, then force a final synthesis. If the tool cannot show its work, it is not doing research. It is doing improvisation.\u003C\u002Fp>\u003Cul>\u003Cli>Require a source list before the final answer.\u003C\u002Fli>\u003Cli>Make the agent separate claims from evidence.\u003C\u002Fli>\u003Cli>Use a second pass to challenge weak or conflicting data.\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>The free part is not a footnote. It changes adoption\u003C\u002Fh2>\u003Cp>The article keeps stressing that Wenxin Assistant is free and unlimited. Usually I would treat that as a side note, but here it matters. Agent products are expensive to run. Tool calls, long contexts, retrieval, document generation, and multi-step planning all cost real money. If a vendor can absorb that cost and keep the product open, that changes how people test it.\u003C\u002Fp>\u003Cp>The practical effect is simple: no one has to do a justification spreadsheet before trying it. That lowers the friction a lot. I have seen teams ignore otherwise decent tools because the first question was pricing, not capability. If the entry point is free, more people will poke at it, break it, and find out whether it actually helps.\u003C\u002Fp>\u003Cp>That does not mean free equals better. It means the product has a distribution advantage, and in agent land, distribution matters because usage exposes failure modes. The more people use a tool on real tasks, the faster its weaknesses show up. If Wenxin is genuinely good at memory, research, and document generation, free access makes that visible faster.\u003C\u002Fp>\u003Cp>How to apply it: if you are shipping an internal agent, think about the adoption tax. Every extra approval, every pricing question, every setup step reduces real usage. Sometimes the best feature is simply removing the reason people hesitate.\u003C\u002Fp>\u003Ch2>The template you can copy\u003C\u002Fh2>\u003Cpre>\u003Ccode># Agent evaluation template: from chat to completed work\n\n## Goal\nBuild an agent that finishes one real task end to end, not just answers questions.\n\n## Task\nUse this agent to complete a multi-step work item:\n- ingest files or sources\n- preserve context across turns\n- call tools when needed\n- produce a final artifact in a usable format\n\n## What to test\n1. Memory\n   - Does the agent remember constraints from earlier turns?\n   - Does it preserve user preferences, scope, and exclusions?\n\n2. Data processing\n   - Does it keep structure intact?\n   - Does it handle tables, fields, dates, and totals without inventing values?\n\n3. Content creation\n   - Does it follow tone, length, and format requirements?\n   - Is the output ready to use without heavy editing?\n\n4. Research analysis\n   - Does it gather sources before concluding?\n   - Does it cross-check conflicting claims?\n   - Does it separate evidence from interpretation?\n\n5. Execution\n   - Does it complete the task without human babysitting?\n   - Does it return a final artifact, not just a plan?\n\n## Scoring\nScore each category from 0 to 5:\n- 0 = failed completely\n- 1 = partial output, major gaps\n- 2 = usable only with heavy edits\n- 3 = mostly correct, some cleanup needed\n- 4 = strong output, minor fixes only\n- 5 = fully usable as delivered\n\n## Pass criteria\nThe agent passes only if:\n- it finishes the task end to end\n- it respects all hard constraints\n- it produces a direct-use artifact\n- it does not require manual reconstruction\n\n## Example prompt\n\"Read the attached files, keep the original constraints in mind, research any missing facts, and return a final [document \u002F report \u002F HTML \u002F spreadsheet] that I can use immediately. Do not ask me to restate requirements. If something is unclear, make the smallest reasonable assumption and note it clearly.\"\n\n## Review checklist\n- [ ] Context preserved across turns\n- [ ] Tools used appropriately\n- [ ] Output format correct\n- [ ] Sources or inputs cited\n- [ ] Final artifact usable without rewrite\n- [ ] No hidden assumptions left unexplained\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>If you want to adapt this to your own team, start by swapping in your actual output type: proposal, report, memo, dashboard, or HTML page. Then add the one failure mode that hurts you most, because that is the test that will tell you whether the agent is real or just talkative.\u003C\u002Fp>\u003Cp>The original source is the Zhihu article at \u003Ca href=\"https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2068369583216537702\">https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2068369583216537702\u003C\u002Fa>. My breakdown is derivative of that post’s claims and examples, but the framing, evaluation template, and implementation advice are my own.\u003C\u002Fp>","Baidu Wenxin’s free agent stack shows how search-engine habits map cleanly onto task execution.","zhuanlan.zhihu.com","https:\u002F\u002Fzhuanlan.zhihu.com\u002Fp\u002F2068369583216537702",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786365227346-puo3.png","tools","en","7c5f393f-e45a-45d0-98fe-28abc173ba11",[17,18,19,20,21],"Baidu Wenxin","AI agents","search","benchmarking","memory",[23,24,25],"Agent quality should be judged by completed tasks, not polished answers.","Search-engine architecture maps naturally onto planning, retrieval, and tool use.","A free agent lowers adoption friction, but only execution quality earns trust.",0,"2026-08-10T12:33:19.92436+00:00","2026-08-10T12:33:19.916+00:00",{"tags":30,"relatedLang":33,"relatedPosts":37},[31],{"name":18,"slug":32},"ai-agents",{"id":15,"slug":34,"title":35,"language":36},"baidu-wenxin-search-to-agent-template-zh","百度文心把搜索底子变成Agent能力","zh",[38,44,50,56,62,68],{"id":39,"slug":40,"title":41,"cover_image":42,"image_url":42,"created_at":43,"category":13},"d5325d38-a48e-43b7-b373-2ecefff808cd","opcode-supports-deepseek-glm-qwen-gpt-models-en","OpenCode now supports DeepSeek, GLM, Qwen, GPT","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786363373903-d4u3.png","2026-08-10T12:02:30.602542+00:00",{"id":45,"slug":46,"title":47,"cover_image":48,"image_url":48,"created_at":49,"category":13},"0aab53f7-2569-4c05-93f1-9bed53def12b","deepseek-codex-ai-coding-costs-en","DeepSeek in Codex Will Cut AI Coding Costs Hard","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786321971893-69cj.png","2026-08-10T00:32:33.297704+00:00",{"id":51,"slug":52,"title":53,"cover_image":54,"image_url":54,"created_at":55,"category":13},"0d0d262b-65b7-4388-82e5-bab51244f9c0","token-vs-word-chinese-tokenization-matters-en","Token vs. word: why Chinese tokenization still matters","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786320173873-ptb3.png","2026-08-10T00:02:32.1747+00:00",{"id":57,"slug":58,"title":59,"cover_image":60,"image_url":60,"created_at":61,"category":13},"0eba42be-cb97-4711-9189-e7f4d0d9ffbe","openai-api-pricing-august-2026-token-costs-en","OpenAI API Pricing Hits $0.05 to $180\u002FM Tokens","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786300371726-ccpx.png","2026-08-09T18:32:26.984366+00:00",{"id":63,"slug":64,"title":65,"cover_image":66,"image_url":66,"created_at":67,"category":13},"d11db4f6-f2c8-4125-8d46-db91fd9c9121","usage-limits-chatgpt-enterprise-edu-controls-en","Usage limits belong in ChatGPT Enterprise and Edu controls","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786298573885-sm6r.png","2026-08-09T18:02:23.482195+00:00",{"id":69,"slug":70,"title":71,"cover_image":72,"image_url":72,"created_at":73,"category":13},"0ad2ef7c-aa71-4608-bd16-e0f27e9dd3de","prepare-for-gemini-3-5-pro-on-launch-day-en","Prepare for Gemini 3.5 Pro on launch day","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786276969432-cdi1.png","2026-08-09T12:02:26.327419+00:00",[75,80,85,90,95,100,105,110,115,120],{"id":76,"slug":77,"title":78,"created_at":79},"8008f1a9-7a00-4bad-88c9-3eedc9c6b4b1","surepath-ai-mcp-policy-controls-en","SurePath AI's New MCP Policy Controls Enhance AI Security","2026-03-26T01:26:52.222015+00:00",{"id":81,"slug":82,"title":83,"created_at":84},"27e39a8f-b65d-4f7b-a875-859e2b210156","mcp-standard-ai-tools-2026-en","MCP Standard in 2026: Integrating AI Tools","2026-03-26T01:27:43.127519+00:00",{"id":86,"slug":87,"title":88,"created_at":89},"165f9a19-c92d-46ba-b3f0-7125f662921d","rag-2026-transforming-enterprise-ai-en","How RAG in 2026 is Transforming Enterprise AI","2026-03-26T01:28:11.485236+00:00",{"id":91,"slug":92,"title":93,"created_at":94},"6a2a8e6e-b956-49d8-be12-cc47bdc132b2","mastering-ai-prompts-2026-guide-en","Mastering AI Prompts: A 2026 Guide for Developers","2026-03-26T01:29:07.835148+00:00",{"id":96,"slug":97,"title":98,"created_at":99},"3ab2c67e-4664-4c67-a013-687a2f605814","garry-tan-open-sources-claude-code-toolkit-en","Garry Tan Open-Sources a Claude Code Toolkit","2026-03-26T08:26:20.245934+00:00",{"id":101,"slug":102,"title":103,"created_at":104},"66a7cbf8-7e76-41d4-9bbf-eaca9761bf69","github-ai-projects-to-watch-in-2026-en","20 GitHub AI Projects to Watch in 2026","2026-03-26T08:28:09.752027+00:00",{"id":106,"slug":107,"title":108,"created_at":109},"9f332fda-eace-448a-a292-2283951eee71","practical-github-guide-learning-ml-2026-en","A Practical GitHub Guide to Learning ML in 2026","2026-03-27T01:16:50.125678+00:00",{"id":111,"slug":112,"title":113,"created_at":114},"1b1f637d-0f4d-42bd-974b-07b53829144d","aiml-2026-student-ai-ml-lab-repo-review-en","AIML-2026 Is a Bare-Bones Student Lab Repo","2026-03-27T01:21:51.661231+00:00",{"id":116,"slug":117,"title":118,"created_at":119},"6d1bf3f6-e191-4d30-b55b-8a0722fa6afe","ai-trending-github-repos-and-research-feeds-en","AI Trending Tracks Repos and Research Feeds","2026-03-27T01:31:35.709532+00:00",{"id":121,"slug":122,"title":123,"created_at":124},"010539a1-4c3a-4bd3-937a-26616422ee0d","awesome-ai-for-science-research-tools-map-en","Awesome AI for Science Is Becoming a Real Research Map","2026-03-27T01:46:50.89513+00:00"]